Homeostatic Noise Buffering in Biomolecular Condensates Hinges on Phase Multiplicity Modulated by Interfacial and Droplet Size Effects

This study investigates how the size of biomolecular condensates and the multiplicity of phases in liquid-liquid phase separation (LLPS) affect the ability of these condensates to buffer against concentration fluctuations (noise). The authors use random phase approximation (RPA) polymer theory and coarse-grained molecular dynamics (MD) simulations to model LLPS of pairs of polyampholytic sequences (model intrinsically disordered proteins, IDPs). The key findings are:

1. RPA predicts both binary and ternary LLPS: "RPA theory predicts both binary and ternary LLPS in a temperature-sensitive manner. We observe demixing underpinned by ternary LLPS for pairs with dissimilar sequence charge patterns but not for pairs with similar sequence charge patterns." The phase behavior is sensitive to the charge patterns of the sequences, with ternary LLPS (three coexisting phases) occurring for some sequence pairs at certain temperatures.

2. MD simulations show finite-size effects: MD-simulated droplets of typical linear dimension ~15 nm deviate significantly from the ideal infinite-volume RPA predictions. the predicted behaviors are corroborated by MD when RPA is augmented with interfacial tension and/or a finite-size formalism commensurating with the typical small sizes of MD model systems. The MD droplets exhibit substantial interfacial features and a high degree of concentration dispersion, which is at variance with conventional infinite-V ternary LLPS.

3. Finite-size effects are captured by phase-field theory and a Gibbs-ensemble-like model: The authors developed a phenomenological phase-field model that extends RPA with an explicit interfacial term. This model, along with a Gibbs-ensemble-like Monte Carlo (MC) model that accounts for thermal fluctuations, successfully rationalizes the diffused density distributions observed in MD. the significant MD concentration dispersions are finite-V interfacial and thermal effects, which will become negligible as V → ∞.

4. Noise buffering depends on droplet size and phase multiplicity: In the infinite-volume limit, ternary LLPS is superior to binary LLPS for noise buffering because the IDP compositions of the three coexisting phases in ternary LLPS remain unchanged over an extended two-dimensional concentration regime, whereas the two coexisting phases in binary LLPS are fixed only along a tieline. However, for small condensates, the buffering capacities of ternary versus binary LLPS are more complex. The analysis concludes that "for droplet linear dimension L ≳ 80 nm, our model polyampholyte ternary LLPS-like systems should behave very similiarly to their corresponding infinite-V ternary systems and thus surpass binary LLPS as a mechanism for concentration noise buffering (Fig. 6)."

5. Implications for MD simulations and biological condensates: The work provides a basis for extrapolating from small-size MD simulations to larger, biologically realistic condensates. The authors note that a model system with ≳ 10 4 molecules is needed for an accurate estimation of the critical temperature Tcr of LLPS and that their estimated threshold of L ≈ 200b ≈ 80 nm for diminished finite-V effects is consistent with this. The findings are relevant for understanding the size diversity of natural biomolecular condensates, from nanocondensates to micron-sized membraneless organelles.

Improvements to AI Systems:

1. Finite-Size-Aware Phase Prediction Models

• Improvement: Integrate the paper’s RPA + interfacial tension + finite-size formalism into machine learning models that predict phase behavior of IDP mixtures. Instead of training on infinite-volume thermodynamic data, augment training with finite-size corrections (e.g., droplet radius, interfacial energy, thermal fluctuation terms) derived from the paper’s phase-field and Gibbs-ensemble models.

• What the improved AI can do: Accurately predict the number and composition of coexisting phases (binary vs. ternary) for condensates of realistic biological sizes (10–200 nm), avoiding the systematic errors that arise when applying infinite-volume predictions to nanoscale droplets.

2. Noise-Buffering Capacity Estimator for Small Condensates

• Improvement: Build a regression or classifier that takes droplet linear dimension (L), sequence charge pattern dissimilarity, and temperature as inputs, and outputs the expected noise-buffering performance (e.g., concentration variance reduction) for binary vs. ternary LLPS. Use the paper’s threshold (L ≈ 80 nm) as a critical feature, and incorporate the finding that ternary LLPS only surpasses binary LLPS above this size.

• What the improved AI can do: Guide experimental design by predicting whether a given IDP pair will act as an effective buffer against concentration fluctuations in vivo, based on the actual size of the condensate, rather than assuming infinite-volume behavior.

3. MD Simulation Extrapolation Tool

• Improvement: Create an AI-based correction layer that takes outputs from small-scale MD simulations (e.g., droplet size 15 nm, high concentration dispersion) and extrapolates them to larger, biologically relevant scales. Train this layer on the paper’s phase-field and Gibbs-ensemble models, which map finite-size MD results to infinite-V predictions.

• What the improved AI can do: Automatically adjust MD-derived phase diagrams and critical temperatures (Tcr) for finite-size artifacts, enabling researchers to estimate the true thermodynamic behavior of condensates without running computationally expensive simulations with >104 molecules.

4. Sequence-Pattern-to-Phase-Behavior Predictor

• Improvement: Use the paper’s finding that ternary LLPS occurs only for pairs with dissimilar sequence charge patterns to train a contrastive or graph neural network that encodes IDP sequences (charge blockiness, net charge, hydropathy) and predicts whether a pair will exhibit binary or ternary LLPS at a given temperature.

• What the improved AI can do: Rapidly screen large libraries of IDP sequences to identify promising pairs for ternary LLPS-based noise buffering, without requiring costly polymer theory calculations or MD simulations.

5. Adaptive Coarse-Graining for Multi-Phase Systems

• Improvement: Develop an AI that learns to switch between infinite-volume thermodynamic models (RPA) and finite-size-corrected models (with interfacial terms) based on the system’s droplet size and the desired accuracy. This could be implemented as a meta-learning module that selects the appropriate physics-based surrogate model.

• What the improved AI can do: Provide real-time, accurate predictions for both nanoscale condensates (where finite-size effects dominate) and micron-sized organelles (where infinite-volume approximations hold), enabling multiscale simulations of cellular phase separation.

6. Concentration-Fluctuation Buffer Design Assistant

• Improvement: Train a generative model (e.g., variational autoencoder) on the paper’s phase diagrams to propose IDP sequence pairs that maximize the concentration regime over which phase compositions remain invariant (the key noise-buffering advantage of ternary LLPS). Constrain the generator with the L > 80 nm requirement for practical buffering.

• What the improved AI can do: Automatically design synthetic IDP pairs for synthetic biology applications, where precise control of intracellular concentration noise is needed, by optimizing for the extended invariant composition regime identified in the paper.

Specific mixing or demixing of molecular species is a characteristic feature of condensed intracellular membraneless compartments. How sequence patterns of intrinsically disordered proteins (IDPs) fundamentally impact subcompartmentalization of biomolecular condensates and their role in buffering against concentration fluctuations are hereby addressed by modeling liquid-liquid phase separation (LLPS) of polyampholytic sequence pairs using random phase approximation (RPA) polymer theory and molecular dynamics (MD). RPA theory predicts both binary and ternary LLPS in a temperature-sensitive manner. We observe demixing underpinned by ternary LLPS for pairs with dissimilar sequence charge patterns but not for pairs with similar sequence charge patterns. Notably, the predicted behaviors are corroborated by MD when RPA is augmented with interfacial tension and/or a finite-size formalism commensurating with the typical small sizes of MD model systems, supporting our stipulation that RPA theory is a useful sequence-specific modeling tool for biomolecular condensates with larger, more realistic sizes when finite-size effects are much less significant. In principle, when the condensate size is sufficiently large, ternary LLPS is superior to binary LLPS in noise buffering because the IDP compositions of the three coexisting phases in ternary LLPS remain unchanged over an extended two-dimensional concentration regime, whereas the two coexisting phases in binary LLPS are fixed only along a tieline. However, when condensate sizes are sufficiently small, the buffering capacities of ternary versus binary LLPSs are more complex as they are modulated differently by finite-size effects. Biophysical ramifications of this interplay are discussed in view of the size diversity of natural biomolecular condensates.